Add --http2-max-concurrent-streams server arg (#34796)

Co-authored-by: Yilong Zhao <74357408+happierpig@users.noreply.github.com>

test/registered/openai_server/basic/test_http2_server.py

@@ -41,7 +41,11 @@ class TestHTTP2Server(CustomTestCase):
             cls.model,
             cls.base_url,
             timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
-            other_args=["--enable-http2"],
+            other_args=[
+                "--enable-http2",
+                "--http2-max-concurrent-streams",
+                "64",
+            ],
         )
 
     @classmethod
@@ -126,7 +130,13 @@ class TestHTTP2ServerMultiTokenizer(TestHTTP2Server):
             cls.model,
             cls.base_url,
             timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
-            other_args=["--enable-http2", "--tokenizer-worker-num", "2"],
+            other_args=[
+                "--enable-http2",
+                "--http2-max-concurrent-streams",
+                "64",
+                "--tokenizer-worker-num",
+                "2",
+            ],
         )
 
 

test/registered/unit/entrypoints/test_http2_server_config.py

@@ -0,0 +1,40 @@
+import importlib.util
+import unittest
+from unittest.mock import patch
+
+from sglang.srt.entrypoints.http_server import _run_granian_server
+from sglang.test.ci.ci_register import register_cpu_ci
+
+register_cpu_ci(est_time=2, suite="base-a-test-cpu")
+
+
+@unittest.skipUnless(
+    importlib.util.find_spec("granian"), "granian is required for HTTP/2"
+)
+class TestGranianHTTP2Config(unittest.TestCase):
+    def test_passes_explicit_max_concurrent_streams(self):
+        configured = {}
+
+        class FakeEmbeddedServer:
+            def __init__(self, **kwargs):
+                configured.update(kwargs)
+
+            async def serve(self):
+                return None
+
+            def stop(self):
+                return None
+
+        with patch("granian.server.embed.Server", FakeEmbeddedServer):
+            _run_granian_server(
+                host="127.0.0.1",
+                port=30000,
+                log_level="info",
+                http2_max_concurrent_streams=37,
+            )
+
+        self.assertEqual(configured["http2_settings"].max_concurrent_streams, 37)
+
+
+if __name__ == "__main__":
+    unittest.main()